xmos-ai-tools 1.3.2.dev80__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/hard_swish.h

@@ -0,0 +1,168 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_HARD_SWISH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_HARD_SWISH_H_
+
+#include <algorithm>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite_micro {
+namespace reference_ops {
+
+inline int16_t SaturatingLeftShift(int16_t value, int amount) {
+  int64_t result = static_cast<int64_t>(value) * (1 << amount);
+  result = std::min<int64_t>(result, std::numeric_limits<int16_t>::max());
+  result = std::max<int64_t>(result, std::numeric_limits<int16_t>::min());
+  return result;
+}
+
+// Similar to ARM instruction SQDMULH.
+// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except
+// rounding to zero instead of to nearest (SQRDMULH).
+inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) {
+  bool overflow = a == b && a == std::numeric_limits<std::int16_t>::min();
+  std::int32_t a_32(a);
+  std::int32_t b_32(b);
+  std::int32_t ab_32 = a_32 * b_32;
+  std::int16_t ab_x2_high16 = static_cast<std::int16_t>((ab_32) / (1 << 15));
+  return overflow ? std::numeric_limits<std::int16_t>::max() : ab_x2_high16;
+}
+
+template <typename T>
+inline void HardSwish(const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float");
+  auto matching_size = MatchingFlatSize(input_shape, output_shape);
+  const T* in_end = input_data + matching_size;
+  for (; input_data < in_end; input_data++, output_data++) {
+    const float in = *input_data;
+    *output_data =
+        in * std::min(static_cast<T>(6), std::max(static_cast<T>(0), in + 3)) /
+        6;
+  }
+}
+
+template <typename T>
+inline void HardSwish(const HardSwishParams& params,
+                      const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Quantized");
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int16_t input_value = input_data[i] - params.input_zero_point;
+    // Left-shift as much as we can without overflow/saturation to put
+    // significant bits in the high bits of our 16-bit fixedpoint values, so
+    // that fixed-point approximate computations below are as accurate as
+    // possible.
+    const int16_t input_value_on_hires_input_scale = input_value * (1 << 7);
+    // Compute the input value on essentially the output scale, just not
+    // right-shifted yet. This is the value that we'll use in the (x >= +3)
+    // case, and that in the general case we'll multiply against the "relu-ish"
+    // fixed-point multiplier in [0, 1].
+    const int16_t input_value_on_preshift_output_scale =
+        gemmlowp::SaturatingRoundingDoublingHighMul(
+            input_value_on_hires_input_scale,
+            params.output_multiplier_fixedpoint_int16);
+    // Now compute the "relu-ish multiplier". In the (-3 <= x <= +3) case, that
+    // is just an affine rescaling of x from [-3, 3] to [0, 1]. In the general
+    // case, it is just that plus saturation at the boundaries of [-3, 3].
+    // First, we rescale from [-3, 3] to [-1, 1], saturating.
+    // That is done by rescaling the input value with a fixed-point multiplier
+    // (reluish_multiplier_fixedpoint) and bit-shift such that we represent
+    // that input value on the scale where the real value 3.0f is represented
+    // by the quantized value 32768.  (+32768 is actually not representable as
+    // int16_t, so this saturates at +32767, and that is seen empirically to be
+    // a negligible contribution to numerical error/bias).
+    //
+    // This code is careful to correctly implement any magnitude of multiplier,
+    // involving either a right shift or a left shift, with correct saturation
+    // behavior in the left-shift case. This forces this code to be more
+    // complicated, but is necessary for real applications: a partially
+    // trained quantized MobileNet v3-small model that motivated this code
+    // exhibits some large [min, max] range boundaries, of the order of
+    // magnitude of 10 or 100 depending on layers.
+    //
+    // The next few lines are basically just an ordinary
+    // MultiplyByQuantizedMultiplier, except that we are more careful here
+    // about the fine details of saturation when left-shifting, because here
+    // overflow in left-shift is a common case, not an anomaly as
+    // MultiplyByQuantizedMultiplier assumes.
+    int16_t reluish_value = input_value_on_hires_input_scale;
+    // Shift left, saturating, as much as we can while ensuring that this
+    // saturation will not contribute to the result. That is, left shift amount
+    // reduced by 1.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(
+          reluish_value, params.reluish_multiplier_exponent - 1);
+    }
+    // Apply the fixed-point multiplier, dividing the value by a divisor
+    // ranging in [1, 2].
+    reluish_value = gemmlowp::SaturatingRoundingDoublingHighMul(
+        reluish_value, params.reluish_multiplier_fixedpoint_int16);
+    // Apply the last bit of left-shift. Thus, in the left-shifting case, if
+    // any saturation affects the result, it is happening here --- any
+    // saturation having occurred above is overwritten here, not affecting the
+    // result.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(reluish_value, 1);
+    }
+    // Shift right, in the right-shifting case.
+    if (params.reluish_multiplier_exponent < 0) {
+      reluish_value = gemmlowp::RoundingDivideByPOT(
+          reluish_value, -params.reluish_multiplier_exponent);
+    }
+    // At this point we have rescaled the value into a 16bit fixedpoint
+    // reluish_value in [-1, 1].
+    // We now convert that to a 16bit fixedpoint value in [0, 1].
+    reluish_value = (reluish_value + (1 << 15)) >> 1;
+    // Use of SaturatingDoublingHighMul here is important to cancel the biases
+    // from the above SaturatingRoundingDoublingHighMul.
+    //
+    // On a partially trained MobileNet-v3-small,
+    //
+    //                                       | bias on    |  ImageNet
+    //                                       | quantized  |  Top-1
+    // Operation used here                   | values     |  accuracy (50k)
+    // --------------------------------------+------------+-----------
+    // SaturatingDoublingHighMul             | -0.0024    |  58.920
+    // SaturatingRoundingDoublingHighMul     | -0.0067    |  58.064
+    //
+    // In activations_test, this is covered by this testcase:
+    //     QuantizedActivationsOpTest.HardSwishBias
+    //
+    const int16_t preshift_output_value = SaturatingDoublingHighMul(
+        reluish_value, input_value_on_preshift_output_scale);
+    // We were so far operating on the pre-shift output scale. Now we finally
+    // apply that output shift, arriving at the final output scale.
+    int16_t output_value = gemmlowp::RoundingDivideByPOT(
+        preshift_output_value, -params.output_multiplier_exponent);
+    output_value += params.output_zero_point;
+    output_value =
+        std::min<int16_t>(output_value, std::numeric_limits<T>::max());
+    output_value =
+        std::max<int16_t>(output_value, std::numeric_limits<T>::min());
+    output_data[i] = output_value;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_HARD_SWISH_H_

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/integer_ops/add.h

@@ -0,0 +1,250 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+
+#include <algorithm>
+#include <cstddef>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite_micro {
+namespace reference_integer_ops {
+
+inline void CheckArithmeticParams(const ArithmeticParams& params) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  // Input offset is negative input zero point. Activation tensors are
+  // asymmetric quantized so they span the full int8 range.
+  TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_GE(-params.input2_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits<int8_t>::max());
+  TFLITE_DCHECK_LE(-params.input2_offset, std::numeric_limits<int8_t>::max());
+}
+
+// TODO: b/270589088 - move to a more appropriate file (b/270589088#comment2)
+template <typename T>
+void BroadcastInput1(int size, const ArithmeticParams& params,
+                     const T* input1_data, const T* input2_data, T* output_data,
+                     void (*check_arithmetic_params)(const ArithmeticParams&),
+                     T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[0], input2_data[i], params);
+  }
+}
+
+template <typename T>
+void BroadcastInput2(int size, const ArithmeticParams& params,
+                     const T* input1_data, const T* input2_data, T* output_data,
+                     void (*check_arithmetic_params)(const ArithmeticParams&),
+                     T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[0], params);
+  }
+}
+
+// TODO: b/270589088 - move to a more appropriate file (b/270589088#comment2)
+template <typename T>
+void ElementWise(int size, const ArithmeticParams& params, const T* input1_data,
+                 const T* input2_data, T* output_data,
+                 void (*check_arithmetic_params)(const ArithmeticParams&),
+                 T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[i], params);
+  }
+}
+
+template <typename T>
+inline void BroadcastAddRecursiveDimensions(
+    const ArithmeticParams& params, int dimension, size_t* input1_offset_p,
+    size_t* input2_offset_p, size_t* output_offset,
+    size_t* compressed_input1_stride, size_t* compressed_input2_stride,
+    size_t* compressed_output_shape, const T* input1_data, const T* input2_data,
+    T* output_data, void (*check_arithmetic_params)(const ArithmeticParams&),
+    T (*binary_func)(T, T, const ArithmeticParams&)) {
+  if (dimension > 0) {
+    for (size_t c = 0; c < compressed_output_shape[dimension]; ++c) {
+      size_t input1_offset_c = *input1_offset_p;
+      size_t input2_offset_c = *input2_offset_p;
+      BroadcastAddRecursiveDimensions(
+          params, dimension - 1, &input1_offset_c, &input2_offset_c,
+          output_offset, compressed_input1_stride, compressed_input2_stride,
+          compressed_output_shape, input1_data, input2_data, output_data,
+          check_arithmetic_params, binary_func);
+      *input1_offset_p += compressed_input1_stride[dimension];
+      *input2_offset_p += compressed_input2_stride[dimension];
+    }
+  } else {
+    TFLITE_DCHECK(dimension == 0);
+    bool input1_is_broadcast = compressed_input1_stride[dimension] == 0;
+    bool input2_is_broadcast = compressed_input2_stride[dimension] == 0;
+    TFLITE_DCHECK(!(input1_is_broadcast && input2_is_broadcast));
+    const T* input1_data_ptr = input1_data + *input1_offset_p;
+    const T* input2_data_ptr = input2_data + *input2_offset_p;
+    T* output_data_ptr = output_data + *output_offset;
+    if (input1_is_broadcast) {
+      // input1 is broadcast.
+      BroadcastInput1<T>(compressed_output_shape[dimension], params,
+                         input1_data_ptr, input2_data_ptr, output_data_ptr,
+                         check_arithmetic_params, binary_func);
+      *input2_offset_p += compressed_output_shape[dimension];
+    } else if (input2_is_broadcast) {
+      // input2 is broadcast.
+      BroadcastInput2<T>(compressed_output_shape[dimension], params,
+                         input1_data_ptr, input2_data_ptr, output_data_ptr,
+                         check_arithmetic_params, binary_func);
+      *input1_offset_p += compressed_output_shape[dimension];
+    } else {
+      // Add element-wise.
+      ElementWise<T>(compressed_output_shape[dimension], params,
+                     input1_data_ptr, input2_data_ptr, output_data_ptr,
+                     check_arithmetic_params, binary_func);
+      *input1_offset_p += compressed_output_shape[dimension];
+      *input2_offset_p += compressed_output_shape[dimension];
+    }
+    *output_offset += compressed_output_shape[dimension];
+  }
+}
+
+// TODO: b/270589088 - move to a more appropriate file. (b/270589088#comment2)
+template <typename T>
+void BroadcastBinaryFunction6DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    T (*binary_func)(T, T, const ArithmeticParams&)) {
+  constexpr int kMaxBroadcastDim = 6;
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  size_t compressed_input1_stride[kMaxBroadcastDim];
+  size_t compressed_input2_stride[kMaxBroadcastDim];
+  size_t compressed_output_shape[kMaxBroadcastDim];
+  bool broadcastable_shape = ReduceDimensionsForBroadcast<kMaxBroadcastDim>(
+      input1_shape, input2_shape, compressed_input1_stride,
+      compressed_input2_stride, compressed_output_shape);
+  // Skip broadcasting for degenerate shapes.
+  if (!broadcastable_shape) {
+    return;
+  }
+
+  size_t input1_offset = 0;
+  size_t input2_offset = 0;
+  size_t output_offset = 0;
+  BroadcastAddRecursiveDimensions(
+      params, kMaxBroadcastDim - 1, &input1_offset, &input2_offset,
+      &output_offset, compressed_input1_stride, compressed_input2_stride,
+      compressed_output_shape, input1_data, input2_data, output_data,
+      check_arithmetic_params, binary_func);
+}
+
+template <typename T>
+void BroadcastBinaryFunction4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    T (*binary_func)(T, T, const ArithmeticParams&)) {
+  BroadcastBinaryFunction6DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                check_arithmetic_params, binary_func);
+}
+
+inline int8_t AddFunc(int8_t x, int8_t y, const ArithmeticParams& params) {
+  const int32_t input1_val = params.input1_offset + x;
+  const int32_t input2_val = params.input2_offset + y;
+  const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+  const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+  const int32_t scaled_input1_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input1_val, params.input1_multiplier, params.input1_shift);
+  const int32_t scaled_input2_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input2_val, params.input2_multiplier, params.input2_shift);
+  const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+  const int32_t raw_output =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          raw_sum, params.output_multiplier, params.output_shift) +
+      params.output_offset;
+  const int32_t clamped_output =
+      std::min(params.quantized_activation_max,
+               std::max(params.quantized_activation_min, raw_output));
+  return static_cast<int8_t>(clamped_output);
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+inline void AddElementwise(int size, const ArithmeticParams& params,
+                           const int8_t* input1_data, const int8_t* input2_data,
+                           int8_t* output_data) {
+  ElementWise(size, params, input1_data, input2_data, output_data,
+              CheckArithmeticParams, AddFunc);
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int8_t* input1_data,
+                const RuntimeShape& input2_shape, const int8_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  CheckArithmeticParams(params);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void BroadcastAdd6DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const int8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const int8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               int8_t* output_data) {
+  BroadcastBinaryFunction6DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                CheckArithmeticParams, AddFunc);
+}
+
+inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const int8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const int8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               int8_t* output_data) {
+  BroadcastBinaryFunction6DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                CheckArithmeticParams, AddFunc);
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

@@ -0,0 +1,241 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+
+#include <algorithm>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite_micro {
+namespace reference_integer_ops {
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  // Get parameters.
+  const int32_t input_offset = params.input_offset;  // r = s(q - Z)
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int32_t output_offset = params.output_offset;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = input_shape.Dims(3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_NE(groups, 0);
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  TFLITE_DCHECK_NE(filters_per_group, 0);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          auto group = out_channel / filters_per_group;
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < filter_input_depth;
+                   ++in_channel) {
+                int32_t input_val =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 32 bits accumulator.
+                // In the nudging process during model quantization, we force
+                // real value of 0.0 be represented by a quantized value. This
+                // guarantees that the input_offset is a int8_t, even though
+                // it is represented using int32_t. int32_t += int8_t *
+                // (int8_t - int8_t) so the highest value we can get from each
+                // accumulation is [-127, 127] * ([-128, 127] -
+                // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                // = 14.98, which means we can accumulate at least 2^16
+                // multiplications without overflow. The accumulator is
+                // applied to a filter so the accumulation logic will hold as
+                // long as the filter size (filter_y * filter_x * in_channel)
+                // does not exceed 2^16, which is the case in all the models
+                // we have seen so far.
+                // TODO(b/174275578): Add a check to make sure the
+                // accumulator depth is smaller than 2^16.
+                acc += filter_val * (input_val + input_offset);
+              }
+            }
+          }
+
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+// 16-bit data and 8-bit filter
+template <typename AccumScalar>
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const AccumScalar* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  // Get parameters.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = input_shape.Dims(3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          auto group = out_channel / filters_per_group;
+          AccumScalar acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < filter_input_depth;
+                   ++in_channel) {
+                int32_t input_val =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 64 bits accumulator.
+                // int64_t += int8_t * int16_t so the highest value we can
+                // get from each accumulation is [-127, 127] * ([-32768,
+                // 32767] -
+                // [-32768, 32767]), which is [-8322945, 8322945].
+                // log2(8322945) = 22.99.
+                acc += filter_val * input_val;
+              }
+            }
+          }
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int16_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_